Seeing the Dark Side: Physics-Conditioned Restoration of Partially Illuminated Objects in Night Images

Preprint, not peer reviewed. At night, a camera often records only the part of an object that a nearby lamp happens to light, while a range sensor still measures the whole shape. We study how much of the dark part can be restored and from which evidence. On a controlled benchmark of 60 real objects under real partial lighting with simulated night-sensor noise, we first measure where a pixel's own signal stops being the best evidence for its appearance: below about 3–10% of the illumination of the lit part without added noise, and below 30% or more under realistic night noise. We then show that residual illumination cannot be estimated reliably from the night image alone, and that Lambertian shading rendered from measured geometry and calibrated light positions lowers the estimation error by 24%; a control run attributes the whole gain to this light prior. On these estimates we build Penumbra, a physics-conditioned restoration framework: the estimated residual illumination, noise level, an SNR-limited relit frame and the light prior condition a restoration network that also predicts a per-pixel gate between the pixel's own signal and the object's appearance, and a calibrated uncertainty. Without any reference information about the dark part, the conditioning lowers the dark-region colour error of the same restoration backbone by 28–30% relative to its RGB-only version and by 22–23% relative to a version given geometry and the light prior as plain image channels; retrained low-light enhancers without it are worse than simple relighting. The gain is independent of the backbone, and a controlled ablation shows that object-level attention adds nothing once residual illumination is estimated. All experiments use real photographs with simulated sensor noise and calibrated light positions; validation on real night captures is ongoing.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24404248
Primary Topic
Image Enhancement Techniques
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Seeing the Dark Side: Physics-Conditioned Restoration of Partially Illuminated Objects in Night Images

Muhammad Djamaluddin, Irwan Rahadi, Siti Hariati Hastuti, Isnein Akbar
1 links
Zenodo (CERN European Organization for Nuclear Research)
Image Enhancement Techniques
preprint

Seeing the Dark Side: Physics-Conditioned Restoration of Partially Illuminated Objects in Night Images

Muhammad Djamaluddin, Irwan Rahadi, Siti Hariati Hastuti, Isnein Akbar
preprint en

Abstract

Preprint, not peer reviewed. At night, a camera often records only the part of an object that a nearby lamp happens to light, while a range sensor still measures the whole shape. We study how much of the dark part can be restored and from which evidence. On a controlled benchmark of 60 real objects under real partial lighting with simulated night-sensor noise, we first measure where a pixel's own signal stops being the best evidence for its appearance: below about 3–10% of the illumination of the lit part without added noise, and below 30% or more under realistic night noise. We then show that residual illumination cannot be estimated reliably from the night image alone, and that Lambertian shading rendered from measured geometry and calibrated light positions lowers the estimation error by 24%; a control run attributes the whole gain to this light prior. On these estimates we build Penumbra, a physics-conditioned restoration framework: the estimated residual illumination, noise level, an SNR-limited relit frame and the light prior condition a restoration network that also predicts a per-pixel gate between the pixel's own signal and the object's appearance, and a calibrated uncertainty. Without any reference information about the dark part, the conditioning lowers the dark-region colour error of the same restoration backbone by 28–30% relative to its RGB-only version and by 22–23% relative to a version given geometry and the light prior as plain image channels; retrained low-light enhancers without it are worse than simple relighting. The gain is independent of the backbone, and a controlled ablation shows that object-level attention adds nothing once residual illumination is estimated. All experiments use real photographs with simulated sensor noise and calibrated light positions; validation on real night captures is ongoing.

Zenodo (CERN European Organization for Nuclear Research)
Universitas Hamzanwadi (ID)
Image Enhancement Techniques
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Seeing the Dark Side: Physics-Conditioned Restoration of Partially Illuminated Objects in Night Images — Muhammad Djamaluddin, Irwan Rahadi, et al. · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS